[Feature] Improve VISReg loss numerical stability and defaults

[HaiyuWu]

• VISRegLoss: clamp std to a minimum instead of adding 1e-6, so the scale term and normalisation stay unbiased for well-conditioned features while still guarding against division-by-zero on collapsed dimensions.
• VISRegLoss: keep the Gaussian quantile target in fp32 rather than casting it to the input dtype, for a more stable shape term under 16-mixed.
• VISReg: change the default convex weight lamb from 0.02 to 0.9.

Description

Checklist

• [ x ] I have read the Contributing document.
• [ x ] The documentation is up-to-date with the changes I made (check build artifacts).
• [ x ] All tests passed, and additional code has been covered with new tests.
• [ x ] I have added the PR to the RELEASES.rst file.